DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark
DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark is a 27 billion parameter Qwen3.6 model, based on a hybrid Mamba2 + Transformer architecture, recovered from a Q8_K_P quantized GGUF by DreamFast. This model has undergone "abliteration" using the Reaper tool, which removes safety filters while largely preserving core capabilities. It is optimized for uncensored content generation and reasoning efficiency, demonstrating strong performance in benchmarks like MMLU and adjusted GSM8K.
Loading preview...
DreamFast/Qwen3.6-27B-Uncensored-HauhauCS-Aggressive-Safetensor-Benchmark Overview
This model is a ~27 billion parameter Qwen3.6 variant, recovered by DreamFast from a Q8_K_P quantized GGUF originally published by HauhauCS. It utilizes a hybrid Mamba2 + Transformer architecture with a 262,144 token context length. The model has been subjected to "abliteration" using the Reaper tool, which aims to remove safety filters without significantly degrading core capabilities.
Key Characteristics & Performance
- Uncensored Output: Achieves near-complete safety removal, with 94.5% reported ASR (Attack Success Rate) reaching 100% with Chain-of-Thought (CoT) analysis on HarmBench.
- Capability Preservation: Demonstrates solid retention of capabilities, with MMLU scores at 83.9% (a +0.6pp increase over the base model) and adjusted GSM8K at 96.6% (only +0.4pp from base).
- Reasoning Efficiency: Abliteration significantly improves reasoning efficiency on tasks like GSM8K by shortening thinking chains, allowing more answers within token budgets, rather than enhancing raw mathematical ability.
- Weight Modifications: Features extensive weight modifications (564 out of 850 language model keys changed), a combination of Reaper's abliteration edits and GGUF quantization round-trip noise. Despite this, KL divergence remains low (0.0242), indicating good output distribution preservation.
Why Choose This Model?
- Unrestricted Content Generation: Ideal for use cases requiring uncensored or aggressive content generation, where safety filters are undesirable.
- Strong Core Performance: Maintains high performance on general knowledge and reasoning tasks, making it suitable for applications where both uncensored output and accuracy are critical.
- Efficiency in Reasoning: Benefits from improved reasoning efficiency, which can lead to faster and more complete responses in complex problem-solving scenarios compared to the base model.